CN107423298B - Searching method and device - Google Patents

Abstract

The present invention provides a search method and device, wherein the method includes: obtaining a current query input by a user; The ranking of the search results obtained by the current query in this search. The invention can make the search result more accurately meet the user's search requirement.

Description

A search method and device

【Technical field】

The present invention relates to the technical field of computer applications, in particular to a search method and device.

【Background technique】

With the rapid development and popularization of computer technology, more and more people use search engines to obtain information. By entering search keywords in the search box, the search engine can return search results matching the search keywords to the user. . However, most of the existing search methods are based on the common needs of users when returning search results to users, and the search results that meet the needs of most users will be ranked higher in the search result page. For example, when a user enters the search keyword "Chen He", most users are looking for Chen He's encyclopedia, so Chen He's encyclopedia page will be ranked at the top.

However, the search needs of most users are not necessarily the search needs of the current user. For example, the last search keyword entered by the user is "Deng Chao Weibo", and the current search keyword is "Chen He", then the user is very There is a high probability that what you are looking for is Chen He's Weibo, not Chen He's Encyclopedia. Therefore, the accuracy of the existing search methods for user search demand positioning needs to be improved.

[Content of the Invention]

In view of this, the present invention provides a search method and device, so that the search results can more accurately meet the user's search requirements.

The specific technical solutions are as follows:

The present invention provides a search method, the method includes:

Get the current query entered by the user;

According to the click probability of each search result corresponding to the current query in the historical search log under the last query condition input by the user, the ranking of each search result obtained by the current query in this search is determined.

According to a preferred embodiment of the present invention, the previous query of the current query is determined in the following manner:

Determine the previous query from the oq field or rq field in the url parameter of the search result page corresponding to the current query; or,

Determine the previous query from the word field of the url parameter in the referer containing the current query's search request.

According to a preferred embodiment of the present invention, according to the click probability of each search result corresponding to the current query in the historical search log under the condition of the previous query input by the user, determining the ranking of each search result obtained by the current query in this search includes: :

Query the click probability model, and determine the click probability of each search result obtained by the current query in the current search under the last query condition, wherein the click probability model is based on the current search log under the previous query condition in the historical search log. The click status of each search result corresponding to the query is obtained by training;

According to the click probability, the ranking of each search result obtained by the current query in this search is determined.

According to a preferred embodiment of the present invention, the click probability model includes:

The relationship between the click probability of each webpage and the characteristics of the webpage;

Wherein, the webpage features include user preference features, or the webpage features include user preference features and non-user preference features.

According to a preferred embodiment of the present invention, the user preference feature includes: a click rate or a click rate of the webpage;

The click-through rate is the ratio of the number of clicks to the number of views of the webpage in the search result page, and the click-through rate is the ratio of the number of clicks to the number of impressions of the webpage in the search result page.

According to a preferred embodiment of the present invention, the non-user preference feature includes: the ranking of the web page in the search result page or the matching degree between the web page and the corresponding query.

According to a preferred embodiment of the present invention, the click probability model is trained in the following manner:

Use the data in the first time period in the historical search log to generate the webpage feature vector;

Using the data in the second time period in the historical search log as a training sample, the click probability model is trained to obtain model parameters.

According to a preferred embodiment of the present invention, the model parameters include weights of webpage feature vectors.

According to a preferred embodiment of the present invention, when using the data in the historical search log in the first time period to generate the webpage feature vector, for each webpage, respectively: determine the user preference characteristics of the webpage under the conditions of the previous query and the current query, and the current User preference characteristics of web pages under query conditions; retain two determined web pages with different user preference characteristics;

The webpage feature vector is generated using the retained webpages.

According to a preferred embodiment of the present invention, the click probability model includes:

Among them, P is the click probability of the web page, x ₁ is the user preference feature vector, x ₂ is the non-user preference feature vector, and θ ₁ and θ ₂ are the weights of x ₁ and x ₂ , respectively.

The present invention also provides a search device, which includes:

Get unit, used to get the current query entered by the user;

The sorting unit is configured to determine the sorting of each search result obtained by the current query in this search according to the click probability of each search result corresponding to the current query in the historical search log under the last query condition input by the user.

According to a preferred embodiment of the present invention, the device further comprises:

The determination unit is used to determine the previous query of the current query in the following ways:

Determine the previous query from the oq field or rq field in the url parameter of the search result page corresponding to the current query; or,

Determine the previous query from the word field of the url parameter in the referer containing the current query's search request.

According to a preferred embodiment of the present invention, the sorting unit specifically includes:

The query subunit is used to query the click probability model, and determine the click probability of each search result obtained by the current query in the current search under the condition of the previous query, wherein the click probability model is based on the historical search log. It is obtained by training the click status of each search result corresponding to the current query under the previous query condition;

The sorting subunit is used for determining the sorting of each search result obtained by the current query in this search according to the click probability.

According to a preferred embodiment of the present invention, the click probability model includes:

The relationship between the click probability of each webpage and the characteristics of the webpage;

Wherein, the webpage features include user preference features, or the webpage features include user preference features and non-user preference features.

According to a preferred embodiment of the present invention, the user preference feature includes: a click rate or a click rate of the webpage;

The click-through rate is the ratio of the number of clicks to the number of views of the webpage in the search result page, and the click-through rate is the ratio of the number of clicks to the number of impressions of the webpage in the search result page.

According to a preferred embodiment of the present invention, the non-user preference feature includes: the ranking of the web page in the search result page or the matching degree between the web page and the corresponding query.

According to a preferred embodiment of the present invention, the device further comprises:

The training unit is used for generating webpage feature vectors by using the data in the first time period in the historical search log; using the data in the second time period in the historical search log as a training sample to train a click probability model to obtain model parameters.

According to a preferred embodiment of the present invention, the model parameters include weights of webpage feature vectors.

According to a preferred embodiment of the present invention, when the training unit uses the data in the historical search log in the first time period to generate the webpage feature vector, for each webpage, respectively execute: determine the user preference of the webpage under the conditions of the previous query and the current query feature, and the user preference feature of the web page under the current query condition, and retain two determined web pages with different user preference features, so as to use the retained web pages to generate the web page feature vector.

According to a preferred embodiment of the present invention, the click probability model includes:

Among them, P is the click probability of the web page, x ₁ is the user preference feature vector, x ₂ is the non-user preference feature vector, and θ ₁ and θ ₂ are the weights of x ₁ and x ₂ , respectively.

It can be seen from the above technical solutions that the present invention determines the order of each search result obtained by the current query in this search according to the click probability of each search result corresponding to the current query under the condition of the previous query in the historical search log, so that the search results are more accurate. Accurately meet the user's search needs.

[Description of drawings]

1 is a flowchart of a method provided by an embodiment of the present invention;

FIG. 2 is a structural diagram of an apparatus provided by an embodiment of the present invention.

【Detailed ways】

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. As used in the embodiments of the present invention and the appended claims, the singular forms "a," "the," and "the" are intended to include the plural forms as well, unless the context clearly dictates otherwise.

It should be understood that the term "and/or" used in this document is only an association relationship to describe the associated objects, indicating that there may be three kinds of relationships, for example, A and/or B, which may indicate that A exists alone, and A and B exist at the same time. B, there are three cases of B alone. In addition, the character "/" in this document generally indicates that the related objects are an "or" relationship.

Depending on the context, the word "if" as used herein can be interpreted as "at" or "when" or "in response to determining" or "in response to detecting." Similarly, the phrases "if determined" or "if detected (the stated condition or event)" can be interpreted as "when determined" or "in response to determining" or "when detected (the stated condition or event)," depending on the context )" or "in response to detection (a stated condition or event)".

The core idea of the present invention is that, after obtaining the current query input by the user, according to the click probability of each search result corresponding to the current query in the historical search log under the condition of the previous query input by the user, determine the current query obtained in this search. The ordering of each search result.

In the embodiment of the present invention, a click probability model can be established according to the historical search log, and by querying the click probability model, the click probability of each search result corresponding to the current query in the historical search log under the previous query condition can be determined. That is to say, the method provided by the present invention may include two stages, one is a model establishment stage, and the other is a model use stage, that is, a stage of realizing the search. However, it should be noted that the establishment of the model and the realization of the search are separated. It can be understood that the model establishment stage is an offline process, the search realization stage is an online process, and the model establishment is periodically and continuously updated. The method provided by the present invention will be described in detail below with reference to the embodiment shown in FIG. 1 .

FIG. 1 is a flow chart of a method provided by an embodiment of the present invention. As shown in FIG. 1 , the method may specifically include the following steps:

In 101, a webpage feature vector is generated using the data in the historical search log in the first time period.

In this embodiment, step 101 and step 102 are the realization of the model establishment stage. For the convenience of understanding, the click probability model involved in the embodiment of the present invention is described first. The click probability model reflects the relationship between the click probability of each web page and the characteristics of the web page. The final effect is that the click probability of each search result corresponding to the current query under the previous query condition can be obtained by querying the click probability model.

The webpage features may include user preference features, which reflect the user's preference for the webpage in the search result page under the conditions of the previous query and the current query. Generally, the higher the user's preference for the webpage, the higher the preference for the webpage the higher the probability of clicks. The spot checking rate or the spotting rate, etc. can be used as the user preference feature.

The so-called check rate refers to the ratio of the number of clicks to the number of views of a web page in the search result page. The number of clicks is easier to understand. If the user clicks on a url in the search results page, the url is clicked once. Whether a url in the search result page is browsed can be reflected by the user's click behavior. It can be determined that the user clicks on the last url in the search result page, and all the urls before this url can be considered to have been browsed once. When determining the checking rate of a url, you can count the number of clicks and views of each url in all search results under the conditions of the previous query and the current query, and then use the ratio of the number of clicks to the number of views to determine the checking rate of each url. Rate.

The so-called click-through rate refers to the ratio of the number of clicks to the number of impressions of a web page in the search results page. As long as a url has been displayed in the search results page, it is considered that the url has been displayed once. When determining the click-through rate of a url, you can count the number of clicks and impressions of each url in all search results under the conditions of the previous query and the current query, and then use the ratio of the number of clicks and the number of impressions to determine the click-through rate of each url. Rate.

In addition, the webpage features may include non-user preference features in addition to the user preference features, so as to weaken the influence of the user preference on the click probability. The non-user preference feature may be the ranking of the web page in the search result page, or the matching degree between the web page and the corresponding query, and the like.

The click probability model can represent the click probability of the webpage through a specific relationship expression (ie, a functional relationship) from the above-mentioned webpage characteristics and model parameters. The model parameter may be the weight of each webpage feature. For example, the click probability can be expressed as P=f(x ₁ , θ ₁ , x ₂ , θ ₂ ), where P is the click probability of the url under the conditions of the previous query and the current query, and x ₁ is the previous query and the current query User preference feature vector under the condition, the user preference feature vector is composed of the user preference features of each url under the previous query and current query conditions, x ₂ is the non-user preference feature vector under the previous query and current query conditions, the non-user preference feature vector The user preference feature vector is composed of the non-user preference features of each url under the conditions of the previous query and the current query, and θ ₁ and θ ₂ are the weights of x ₁ and x ₂ , respectively. The functional relationship f() can be a linear relationship or a nonlinear relationship. As one of the implementation methods, the following functional relationship can be used:

In this step, the number of clicks and browsing times of each url can be counted according to the conditions of the previous query and the current query (here, the click-through rate is taken as the user preference feature as an example, if the click-through rate is taken as the user preference feature, then Count the number of impressions, and follow-up processing is similar). In order to improve the efficiency, only the urls with the number of browsing times greater than or equal to the preset number of browsing times can be reserved for the construction of the feature vector of the web page. Because the number of browsing times is very low, the user is less interested in them, so there is no need to do so. for model building. For example, if the first time period is 17 days before the day before the current date, the number of clicks and browsing times of each url under the previous query and the current query will be counted from the historical search log of the 17 days. The URLs with the number of views greater than or equal to 20 are retained, and other URLs are filtered out.

Then calculate the check rate of each url under the conditions of the previous query and the current query. Further, each url can be sorted according to the checking rate, and then the checking rate of each url under the current query condition is calculated, and the sorting is performed according to the checking rate. If the position of a url is the same in the two sorts, it means that the previous query has no effect on the current query on this url, so this url can also be filtered out. The so-called urls under the current query conditions refer to the limitations of the previous query, including all previous queries and no previous query. This process can be summarized as: determining the user preference characteristics of the url under the previous query and the current query condition, as well as the user preference characteristics of the url under the current query condition, and retaining the two determined urls with different user preference characteristics. Subsequently, the webpage feature vector is generated by using the reserved url.

So far, using the reserved url can generate the vector of the check rate, that is, the user preference feature vector. It is also possible to further count the ranking of the retained URLs (it may also be the degree of matching between the URLs and the corresponding query), that is, to determine the non-user preference feature vector of the retained URLs.

In addition, in this step, when determining the previous query of the current query, the following methods may be adopted but not limited to:

The first method: Determine the previous query from the oq field of the url parameter of the search result page corresponding to the current query.

If the user enters a new query2 in the search box after inputting query1 in the search box, the information of query1 will be carried in the oq field of the url parameter of the search result page corresponding to query2, then the upper A query is query1.

If the user clicks a recommended resource (such as a recommended online application) in the search result page of query1 after inputting query1 in the search box, then on the search result page corresponding to the recommended resource (the name of the recommended resource is query2) The oq field of the url parameter will also carry the information of the query1, then the previous query of query2 is query1.

The second method: Determine the previous query from the rq field in the url parameter of the search result page corresponding to the current query.

If the user clicks the related search query2 in the search result page of query1 after entering query1 in the search box, the information of query1 will be carried in the rq field of the url parameter of the search result page of query2 that is jumped to, then query2 The previous query is query1.

The third method: Determine the previous query from the word field of the url parameter in the referer of the search request containing the current query.

Usually, when a browser requests a page, it will include a referer in the request to indicate which web page it is linked from. Of course, the premise is that the url of the referer is the main domain that provides the search service, for example, the main domain is "baidu.com".

In 102, the click probability model is trained by using the data in the second time period in the historical search log as a training sample to obtain model parameters.

The second time period involved in this step is not necessarily related to the first time period, and the two may or may not overlap with each other. For example, the web page feature vector is generated by taking the 17 days before the day before the current date as the first time period, and the data on the day before the current date is used as the training sample to train the click probability model.

In the training sample, if a url is clicked in the search result corresponding to the current query under the previous query condition, the click probability of the url is 1; if it is not clicked, the click probability of the url is 0. Generated feature vectors for training. Assuming that the click probability model is P=f(x ₁ , θ ₁ , x ₂ , θ ₂ ), after using the generated x ₁ and x ₂ and training data for training, the weights θ ₁ and 2 of x ₁ and x ₂ can be obtained. θ ₂ .

To understand the click probability model more clearly, take an example:

Statistics are performed on the data in the first time period in the historical search log, and the data shown in Table 1 can be obtained. Since there are many data obtained by statistics, only a part of them is intercepted in Table 1.

Table 1

In Table 1, each web page in the search result can be identified by a url or a resource identifier. It can be seen from Table 1 that the current query is also "Anthony". Due to the difference of the previous query, the situation where the same web page is clicked or browsed by the user is different. When it was "Wade", it was viewed 178 times and clicked 29 times; when the last query was "Spend a long time with Anthony" (this is the name of a novel), it was viewed 2429 times and was clicked 2286 times. Obviously, when the user first enters "Accompany Anthony through the long years" and then enters "Anthony", the probability of clicking on the page is higher. When this happens in the future, it should be ranked higher in the search results. .

After training the data in the second time period as a training sample, the final click probability model can contain the data shown in Table 2 below:

Table 2

In Table 2, the number before ":" is the page feature value, and the number after ":" is the page feature identifier. For example, "72", "26" and "71" in Table 2 are used to identify different The page features are three page features, namely, "discrete value of spot inspection rate_rank in search results", "rank in search results" and "discrete value of spot inspection rate". The discrete value of the inspection rate refers to a value formed after the inspection rate is discretized, for example, the inspection rate is discretized into an integer value between 0 and 20, and of course, it may not be discretized. In addition, since "rank in search results" is a non-user preference feature, this feature is to weaken the influence of user preference feature on click probability, so it can be seen from Table 2 that its weight value is negative.

So far, the click probability model has been established. For the combined conditions of the previous query and the current query, each url has a corresponding click probability. The following steps are the online phase to implement the search.

In 103, the current query input by the user is acquired, and the current query is used to perform search matching to obtain each search result.

In 104, the click probability model is queried, and the click probability of each search result obtained by the current query in the current search under the condition of the previous query is determined.

Assuming that the user currently enters query2, and the last input query is query1, then query the click probability model to determine the click probability of each search result obtained by this search of query2 under the condition that query1 is the previous query of query2.

In 105, according to the click probability, the ranking of each search result obtained by the current query is determined.

In this step, the click probability can be used as one of the factors for sorting the search results. For example, if the original search results are sorted according to the text similarity between each search result and the current query, the click probability factor can be added. , the text similarity and the click probability are used such as weighting to determine the ranking score of each search result, and then the ranking is performed according to the ranking score. In the case of the same text similarity, the search results with higher click probability will be ranked higher.

Alternatively, each search result in the originally obtained search result page may be correspondingly increased or decreased in the order of the search result page according to the click probability. For example, the click probability can be divided into several levels, for example, those with a click probability higher than 80% belong to the first level, those with a click probability of 60% to 80% belong to the second level, those with a click probability of 30% to 60% belong to the third level, and those with a click probability of 30% to 60% belong to the third level. 30% belong to the fourth level, and then the ranking of the URLs with the click probability at the first level is increased by n1, and the ranking of the URLs at the second level is increased by n2. If n1>n2, the URLs at the third level are ranked The ranking remains unchanged, and the ranking of the url at the fourth level is reduced by n3 places.

The above is a detailed description of the method provided by the present invention, and the device provided by the present invention is described in detail below with reference to the embodiments.

FIG. 2 is a structural diagram of an apparatus provided by an embodiment of the present invention. The apparatus may be set on a server side that provides a search service. As shown in FIG. 2, the apparatus may include: an obtaining unit 10 and a sorting unit 20, and may further include a determining unit 30 and training unit 40.

The obtaining unit 10 is responsible for obtaining the current query input by the user. The search engine on the server side will use the current query to search in the search database, and the search matching method is not limited in the present invention.

The sorting unit 20 is responsible for determining the sorting of each search result obtained by the current query in this search according to the click probability of each search result corresponding to the current query in the historical search log under the last query condition input by the user.

The determining unit 30 is responsible for determining the previous query of the current query, which may be in but not limited to the following ways:

The first method: Determine the previous query from the oq field of the url parameter of the search result page corresponding to the current query.

The second method: Determine the previous query from the rq field in the url parameter of the search result page corresponding to the current query.

The third method: Determine the previous query from the word field of the url parameter in the referer of the search request containing the current query.

The sorting unit 20 may specifically include: a query subunit 21 and a sorting subunit 22 .

The query sub-unit 21 is responsible for querying the click probability model, and determining the click probability of each search result obtained by the current query in this search under the previous query condition, wherein the click probability model is based on the historical search log corresponding to the current query under the previous query condition. The click status of each search result is obtained by training.

The click probability model includes: the relationship between the click probability of each webpage and the characteristics of the webpage. The webpage features include user preference features, or the webpage features include user preference features and non-user preference features.

The user preference feature may include: the click rate or the click rate of the web page, and the like. The click-through rate is the ratio of the number of clicks to the number of views of the webpage in the search result page, and the click-through rate is the ratio of the number of clicks to the number of impressions of the webpage in the search result page. The non-user preference features include: the ranking of the web page in the search result page or the degree of matching between the web page and the corresponding query, and the like. For the specific description of the foregoing features, reference may be made to the foregoing method embodiments, which will not be repeated here.

The click probability model can represent the click probability of the webpage through a specific relationship expression (ie, a functional relationship) from the above-mentioned webpage characteristics and model parameters. The model parameter may be the weight of each webpage feature. For example, the click probability can be expressed as P=f(x ₁ , θ ₁ , x ₂ , θ ₂ ), where P is the click probability of the url under the conditions of the previous query and the current query, and x ₁ is the previous query and the current query. User preference feature vector under the condition, the user preference feature vector is composed of the user preference features of each url under the previous query and current query conditions, x ₂ is the non-user preference feature vector under the previous query and current query conditions, the non-user preference feature vector The user preference feature vector is composed of the non-user preference features of each url under the previous query and current query conditions, and θ ₁ and θ ₂ are the weights of x ₁ and x ₂ , respectively. The functional relationship f() can be a linear relationship or a nonlinear relationship. As one of the implementation methods, the following functional relationship can be used:

The sorting subunit 22 is responsible for determining the sorting of each search result obtained by the current query in this search according to the click probability. Specifically, the click probability can be used as one of the factors for sorting the search results. For example, if the original search results are sorted according to the text similarity between each search result and the current query, then the click probability factor can be added, and the The text similarity and click probability determine the ranking score of each search result in a way such as weighting, and then sort according to the ranking score. In the case of the same text similarity, the search results with higher click probability will be ranked higher. Alternatively, each search result in the originally obtained search result page may be correspondingly increased or decreased in the order of the search result page according to the click probability.

The above click probability model is obtained by training by the training unit 40 . The training unit 40 can use the data in the first time period in the historical search log to generate the webpage feature vector; use the data in the second time period in the historical search log as a training sample to train the click probability model to obtain model parameters. There is no necessary relationship between the second time period and the first time period, and the two may or may not overlap each other. For example, the web page feature vector is generated by taking the 17 days before the day before the current date as the first time period, and the data on the day before the current date is used as the training sample to train the click probability model.

The determination of the last query in the historical search log can also be implemented by the determination unit 30 and provided to the training unit 40 .

When the training unit 40 uses the data in the historical search log in the first time period to generate the webpage feature vector, it can respectively perform for each webpage: determine the user preference characteristics of the webpage under the previous query and the current query condition, and the webpage under the current query condition. The user preference features are retained, and the two determined web pages with different user preference features are retained, so as to use the retained web pages to generate a web page feature vector. For example, each url can be sorted according to the checking rate, and then the checking rate of each url under the current query condition is calculated, and the sorting is performed according to the checking rate. If the position of a url is the same in the two sorts, it means that the previous query has no effect on the current query on this url, so this url can also be filtered out. The so-called urls under the current query conditions refer to the limitations of the previous query, including all previous queries and no previous query.

For the effects produced by the above-mentioned embodiments of the present invention, here are a few examples:

If the last query entered by the user is "Wade", and the current query entered is "Anthony", because in the historical search log, when the user enters "Wade" and then enters "Anthony", click Anthony as the search result. There is a high probability of urls for basketball players, so these urls are ranked high in the search results for this search.

If the previous query entered by the user is "Spend the long years with Anthony", and the current query entered is "Anthony", because in the historical search log, when the user enters "Spend the long years with Anthony" and then enters "Anthony" , there is a high probability of hitting writer-related urls in the search results, so these urls are ranked higher in the search results for this search.

If the last query entered by the user is "Faye Wong" and the current query entered is "Red Bean", because in the historical search log, when the user enters "Faye Wong" and then enters "Red Bean", click on the red bean in the search results as a song The probability of urls is very high, so these urls are ranked higher in the search results of this search.

If the previous query entered by the user is "black bean" and the current query entered is "red bean", because in the historical search log, when the user enters "black bean" and then enters "red bean", the user will click on the red bean as a food in the search results The probability of urls is very high, so these urls are ranked higher in the search results of this search.

If the last query entered by the user is "Deng Chao Weibo", and the current query entered is "Chen He", because in the historical search log, when the user enters "Deng Chao Weibo" and then "Chen He", the search results will The probability of clicking on the url of Chen He Weibo is very high, so the url can be ranked high in the search results of this search.

If the last query entered by the user is "Deng Chao", and the current query entered is "Chen He", because in the historical search log, when the user enters "Deng Chao" and then enters "Chen He", click Chen He and Chen He in the search results. The probability of running man-related urls is high, so the url can be ranked high in the search results of this search.

It can be seen that even if the previous query input by the user and the current query are not semantically related, the relationship between the two can be found, and the user's preference can be reflected in the search results. And when there are multiple semantics for the current query, the user's needs can be well distinguished based on the previous query.

In the several embodiments provided by the present invention, it should be understood that the disclosed apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there may be other division manners in actual implementation.

The units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware, or may be implemented in the form of hardware plus software functional units.

The above-mentioned integrated units implemented in the form of software functional units can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute the methods described in the various embodiments of the present invention. some steps. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes .

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the present invention. within the scope of protection.

Claims (20)

1. A method of searching, the method comprising:

acquiring a current query input by a user;

and determining the sequence of each search result obtained by the current query in the current search according to the click probability of each corresponding search result of the current query in the historical search log under the previous query condition input by the user.

2. The method of claim 1, wherein the last query of the current query is determined by:

determining the last query from oq fields or rq fields in url parameters of a search result page corresponding to the current query; or,

and determining the last query from the word field of url parameters in the referrer containing the search request of the current query.

3. The method of claim 1, wherein determining the ranking of the search results obtained by the current query in the current search according to the click probability of the search results corresponding to the current query in the historical search log under the previous query condition input by the user comprises:

inquiring a click probability model, and determining the click probability of each search result obtained by the current query under the previous query condition in the current search, wherein the click probability model is obtained by utilizing the click condition of each search result corresponding to the current query under the previous query condition in the historical search logs in a training mode;

and determining the sequence of each search result obtained by the current query in the current search according to the click probability.

4. The method of claim 3, wherein the click probability model comprises:

the click probability of each web page and the web page characteristics;

wherein the web page features comprise user preference features or the web page features comprise user preference features and non-user preference features.

5. The method of claim 4, wherein the user preference features comprise: the click rate or click-to-display rate of the webpage;

the click rate is the ratio of the number of clicks of the web page in the search result page to the number of browsing times, and the click rate is the ratio of the number of clicks of the web page in the search result page to the number of showing times.

6. The method of claim 4, wherein the non-user-preferred features comprise: ranking of the web pages in the search result page or the matching degree between the web pages and the corresponding query.

7. The method of claim 4, wherein the click probability model is trained by:

generating a webpage feature vector by using data in a first time period in the historical search log;

and training a click probability model by using data in a second time period in the historical search logs as training samples to obtain model parameters.

8. The method of claim 7, wherein the model parameters comprise weights of web page feature vectors.

9. The method of claim 7, wherein when generating the webpage feature vector by using the data in the history search log in the first time period, respectively performing for each webpage: determining user preference characteristics of a webpage under a previous query and a current query condition and user preference characteristics of the webpage under the current query condition; reserving the two determined webpages with different user preference characteristics;

and generating the webpage feature vector by using each reserved webpage.

10. The method of claim 4, wherein the click probability model comprises:

wherein P is the click probability of the web page, x₁Preference feature vector, x, for user₂For non-user-preferred feature vectors, θ₁And theta₂Are respectively x₁And x₂The weight of (c).

11. A search apparatus, characterized in that the apparatus comprises:

the acquisition unit is used for acquiring the current query input by the user;

and the sorting unit is used for determining the sorting of each search result obtained by the current query in the current search according to the click probability of each corresponding search result of the current query in the historical search log under the last query condition input by the user.

12. The apparatus of claim 11, further comprising:

a determining unit, configured to determine a previous query of the current query by using the following method:

determining the last query from oq fields or rq fields in url parameters of a search result page corresponding to the current query; or,

and determining the last query from the word field of url parameters in the referrer containing the search request of the current query.

13. The apparatus according to claim 11, wherein the sorting unit specifically includes:

the query subunit is configured to query a click probability model, and determine click probabilities of search results obtained by the current query in the current search under the previous query condition, where the click probability model is obtained by using click conditions of search results corresponding to the current query under the previous query condition in a historical search log and training the click conditions;

and the sequencing subunit is used for determining the sequencing of each search result obtained by the current query in the current search according to the click probability.

14. The apparatus of claim 13, wherein the click probability model comprises:

the click probability of each web page and the web page characteristics;

wherein the web page features comprise user preference features or the web page features comprise user preference features and non-user preference features.

15. The apparatus of claim 14, wherein the user preference feature comprises: the click rate or click-to-display rate of the webpage;

the click rate is the ratio of the number of clicks of the web page in the search result page to the number of browsing times, and the click rate is the ratio of the number of clicks of the web page in the search result page to the number of showing times.

16. The apparatus of claim 14, wherein the non-user-preferred features comprise: ranking of the web pages in the search result page or the matching degree between the web pages and the corresponding query.

17. The apparatus of claim 14, further comprising:

the training unit is used for generating a webpage feature vector by using data in a first time period in the historical search log; and training a click probability model by using data in a second time period in the historical search logs as training samples to obtain model parameters.

18. The apparatus of claim 17, wherein the model parameters comprise weights of web page feature vectors.

19. The apparatus of claim 17, wherein the training unit, when generating the feature vector of the web page using the data in the historical search log in the first time period, performs, for each web page: determining user preference characteristics of the web pages under the previous query and the current query and user preference characteristics of the web pages under the current query, reserving the two determined web pages with different user preference characteristics, and generating the web page characteristic vector by utilizing the reserved web pages.

20. The apparatus of claim 14, wherein the click probability model comprises:

wherein P is the click probability of the web page, x₁Preference feature vector, x, for user₂For non-user-preferred feature vectors, θ₁And theta₂Are respectively x₁And x₂The weight of (c).

Images